cmp50hx-unlock

A/B: stock driver vs stock driver + EFI unlock (2026-09-10, host .224)

Question: does the EFI unlock alone give a completely untouched system (stock 610.43.03 open module, zero patches) full compute speed?

Method

Matrix

Cell Module EFI FP64 FP32 FP16 half2 INT8 DP4A Coalesc R/W PCIe
A patched no 0.463 14.874 29.390 48.169 540.6/509.5 1.59/1.71 (Gen2 x4)
B patched yes 14.866 29.171 48.056 ~same 1.59/1.70 (Gen2 x4)
C stock no 0.423 0.427 26.923 1.689 504.4/471.8 0.79/0.85 (Gen1)
D stock yes 0.420 13.512 26.698 47.984 503.8/472.0 0.79/0.86 (Gen1)

(TFLOP/s for FP*, TIOP/s for INT8, GB/s for memory/PCIe.)

Results

  1. Cell D is the headline: with zero kernel patches, the EFI unlock alone restores full compute — FP32 0.427 → 13.512 (31.7x), DP4A 1.689 → 47.984 (28.4x). The stock driver booted the pre-unlocked state cleanly: no Xid, no GSP RPC timeout (the masked-timeout path in patch 01 is not needed when the state arrives pre-unlocked pre-OS).
  2. Cells A vs B: the EFI path is performance-neutral on top of the patched module (deltas within run noise).
  3. The 50HX lock profile on this driver: FFMA-FP32 and DP4A are clamped (~1/32, ~1/28); FP16 half2 FMA and INT32 are NOT clamped; memory bandwidth is not clamped.
  4. Out of EFI scope, as designed: PCIe stays Gen1 without kernel patch 04 (cells C/D), ReBAR and RT-count likewise need the module.
  5. FP32 is ~10% lower with the stock module (13.5) than the patched module (14.9); the patched-module runs also carry Gen2. Likely clock/boost behavior under the module’s flow — not investigated further.

Rollback: patched module restored from ~/cmp50hx-all-patched-backup-1029, verified by hash, dmesg marker (21 lines), and a clean P8 idle (zero Xid).